Tencent Hy4: 770B Open-Source Model at $0.90/M — Vision, Computer Use, Weights Pending

Tencent shipped the full Hy4 on November 4, 2026, and the headline is not the parameter count. It is the first model in the Hunyuan line that can see images and operate a computer: opening a command line, clicking through a graphical interface, and working inside professional software like Blender and CAD. Tencent calls it one of the world's leading open-source AI models, and prices it like a budget flagship — about $0.90 per million input tokens and $2.70 per million output tokens.

Read the announcement closely and the "open-source" claim carries an asterisk. The weights are not out yet. Tencent says it "plans to release model weights and system cards in the coming days," which means that, on launch day, every Hy4 request runs through Tencent's own servers or its API partners. Here is what Hy4 actually is, what the pricing fine print says, and what Tencent's own benchmarks do — and do not — prove.

What Hy4 actually is

Hy4 is Tencent's latest-generation Hunyuan large language model, the successor to the Hy4 preview released August 28, 2026. The architecture is a mixture-of-experts (MoE) design: independent reporting puts the scale at 770 billion total parameters with 49 billion active per token, up from 295 billion in Hy3. One caution here: those parameter figures come from the preview's material, and the November 4 announcement does not restate them, a discrepancy flagged by independent coverage of the launch. The context window sits at roughly one million tokens, in the same class as the largest frontier models.

Three things define the release. First, visual understanding: Hy4 can process fine-grained images, analyze complex charts, and do mathematical visual reasoning. Second, computer use: it can interact with command-line interfaces and graphical user interfaces, and Tencent says it was trained on professional software including Blender and CAD to plan and execute long-horizon tasks. Third, continuity on price: Tencent kept the API pricing unchanged from the preview despite what it describes as significant performance gains.

The model is available now through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI assistant Dayuan, and ima. Third-party API access is being set up through OpenRouter, Hermes, and OpenCode, per the announcement. Where the weights themselves stand is the key open question (see the catches section below).

The pricing: $0.90 in, $2.70 out, 50% off through December 31

Tencent's announcement states the rates directly: "approximately US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens." That is the full-model list price, and the cached-input rate (95% below the input rate) is the number that matters most for agent workflows that re-read long contexts.

Two dated promos change the math for early users. Hy4 runs at 50% off on Tencent Cloud TokenHub through December 31, 2026, which drops the effective rates to $0.45 per million input tokens and $1.35 per million output tokens. And WorkBuddy and CodeBuddy, Tencent's workplace and coding assistants, are free from November 4 to 15, so trying Hy4's capabilities inside real products costs nothing for the first 12 days.

For readers in the UK and Canada, at November 5, 2026 market rates (verified at publication: 1 USD = 0.742 GBP / 1.394 CAD), the standard rates work out to roughly £0.67 / C$1.25 per million input tokens, £2.00 / C$3.77 per million output tokens, and £0.033 / C$0.063 per million cached tokens. Under the TokenHub promo, those become roughly £0.33 / C$0.63 for input and £1.00 / C$1.88 for output. The Chinese-language announcement prices the same grid at 6 yuan per million input tokens and 18 yuan per million output tokens, consistent with Tencent's "approximately" USD figures.

The forward-looking part of the pricing story is Hy4 Flash, "scheduled for release soon" according to Tencent, positioned as a lighter model with greater cost-efficiency. No specs, no date, and no price were given, so budget planning should treat it as unannounced until Tencent publishes a rate card.

How the price stacks up

Against the budget flagships already covered on this site, Hy4 sits squarely in the middle of the open-weight pack:

Hy4 is cheaper than Kimi K3.1 Pro and above DeepSeek's off-peak rates, though DeepSeek undercuts it by 6x when you can run off-peak. The TokenHub 50% discount closes the gap with LongCat-2.5's promo pricing almost exactly. What none of the rivals offer on paper yet is Tencent's combination of vision plus computer use at a sub-$1 input rate.

Tencent's benchmarks, and why they need labels

Tencent published several evaluation claims. Every one of them comes from Tencent's own testing, with no independent reproduction available yet. That matters more than usual here because the headline figures are unusually specific.

The strongest claim is a finance evaluation: in a blind assessment run with Tencent Financial Technology across 579 real-world financial tasks, scored by 213 financial professionals, Hy4 achieved an overall performance score of 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017, 2.68 for Kimi K3.1 Pro, and 2.68 for GLM-5.3. Tencent says this made Hy4 the top-performing open-source model in that test. The internal Investment Research Agent Benchmark (iRaB), run by financial AI company Rabyte, showed a 12.03 percentage-point improvement over the Hy4 preview.

On hallucinations, Tencent's internal evaluation reports the rate fell from 12.95% in the Hy4 preview to 7.2%, while errors caused by unsupported assumptions dropped from 23.1% to 5.6%. On WorkBuddy itself, Tencent reports an average performance score of 83.18 and a 96.15% pass rate, with Hy4 ranking first in nine out of twelve scenarios within the mainland Chinese market. Game development benchmarks improved 19.1% (GameDev) and 9.8% (GameLogic) over the preview.

Treat all of these as Tencent's own numbers until independent labs run the model. The company itself was candid about the bigger picture: its announcement acknowledges a "large gap" between Hy4 and the top closed models, according to independent translation of the release. That honesty is unusual in launch announcements; it reframes the release as a race for the open-weight crown rather than the frontier.

The catches: what you cannot do with Hy4 yet

The open-weights story is incomplete. Tencent "plans to release model weights and system cards in the coming days," but on November 4 the weights were not downloadable, the license for the final model was not stated (the preview shipped under Apache 2.0), and no date was given. Independent coverage confirmed that the "model weights" button on the OpenRouter listing points to a Hugging Face URL that is not publicly accessible. Until the weights land, Hy4 cannot be self-hosted: every request goes through Tencent's API.

The safety disclosures are also company-authored. Tencent says offensive cybersecurity materials were excluded from training, and that its Xuanwu and Keen security labs found "no significant concerns" about Hy4's offensive capabilities in the areas evaluated. On biosafety, the announcement says the model's open-ended experimental design capabilities are "more limited compared to frontier models." These are statements of intent and internal testing, not independent audits; the system card, when published, should carry the fuller evaluations.

Finally, the "joining the ranks of the world's leading open-source AI models" framing deserves the label Tencent itself gives it: company marketing. Without published weights, an independent benchmark suite, or a license, Hy4 is best described as a hosted flagship with an open-weights promise, one that the coming days will either confirm or quietly stretch.

Pro tips and common mistakes

If you are budgeting Hy4 into a pipeline, three moves matter. First, use the TokenHub 50% window: December 31 is a hard date, so test workloads now and lock in cost data before list pricing returns — and budget for the return, since promo rates are temporary by definition. Second, design around the cached-input rate: at $0.045 per million, re-reading a million-token context is nearly free, which favors agentic loops that revisit documents over ones that re-prompt from scratch. Third, treat the free WorkBuddy/CodeBuddy window (November 4–15) as your hands-on evaluation period: it is the cheapest way to see whether the computer-use claims survive contact with your actual software.

The common mistake is reading the price as the story. The real question is whether vision plus computer use at $0.90 per million input tokens does something for your workflow that a cheaper text model plus separate tooling does not. If your tasks are text-only, DeepSeek's off-peak rates are 6x cheaper; if they involve charts, GUIs, and long-horizon software work, Hy4 is the first sub-$1 input open-weights option with both capabilities. Also note the parallel on Tencent's enterprise side: Tencent is not the only Asian tech giant packaging AI as metered token products — KT's Token Factory is playing the same marketplace game in Korea, so expect more of this pricing model across the region.

Is Hy4 really open source?

Not yet. Tencent released the full model on November 4, 2026 with a promise to publish the weights and system cards "in the coming days." Until then it is a hosted model accessed through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, and API partners. The Hy4 preview from August 2026 did ship under Apache 2.0, but the final model's license has not been stated.

How much does Hy4 cost?

About $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens, roughly £0.67 / C$1.25, £2.00 / C$3.77, and £0.033 / C$0.063 at November 5, 2026 market rates. A 50% launch discount on Tencent Cloud TokenHub runs through December 31, 2026.

What can Hy4 do that the preview could not?

Visual understanding and computer use, both firsts for the Hunyuan line. Hy4 can analyze images and charts, operate command-line and graphical interfaces, and was trained on professional software including Blender and CAD for long-horizon tasks.

Is the 50% discount real or a promo?

It is a genuine promo with a hard end date: 50% off standard pricing on Tencent Cloud TokenHub through December 31, 2026, after which list rates apply. Separately, WorkBuddy and CodeBuddy are free from November 4 to 15, 2026.

How does Hy4 compare to DeepSeek and Kimi?

In Tencent's own financial evaluation, Hy4 scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for Kimi K3.1 Pro, but that is Tencent's test, not an independent benchmark. On price, Hy4 ($0.90/$2.70) sits between DeepSeek's off-peak rates ($0.15/$0.60) and Kimi K3.1 Pro ($2/$10).

Sources

• Tencent official announcement: Tencent Launches Hy4 (November 4, 2026 — pricing, promos, benchmarks, weights promise)
• TechNode: Tencent launches Hy4 with vision and computer-use capabilities (November 4, 2026 — USD price conversion, free WorkBuddy window)
• Tech-insider: Tencent Hy4: 770B Params, 1M+ Token Context (November 4, 2026 — 770B/49B specs, SCMP open-weight-race framing)
• Maccus.fr: Tencent says Hy4 ranks among the best open models. Its weights aren't out yet (November 4, 2026 — CNY pricing grid, gap-to-closed-models admission, OpenRouter/HF weights status)
• Dev.to: Tencent's Hy4 Can See and Use Your Computer — Explained Honestly (November 4, 2026 — availability surfaces, what's-not-yet framing)
• Emergent.sh: Tencent Hy4 Launch: 770B Vision Model With Computer Use (November 4, 2026 — launch summary, 50% discount)
• The Daily Flare: Tencent launches Hy4 with visual understanding and computer-use capabilities (November 4, 2026 — Blender/CAD training, availability)
• UnicornAPI: Tencent releases full Hy4 AI model (November 4, 2026 — preview-to-launch changes, OrcaRouter spec analysis)
• Selnoviktech: Tencent Officially Launches Hy4 (November 4, 2026 — MoE architecture, weights timeline)